(synopsis is probably incomplete, please change as you see fit)

Originally reported by Tom Deneau on jmh-dev, and then on hotspot-runtime-dev:
  http://mail.openjdk.java.net/pipermail/hotspot-runtime-dev/2014-November/013216.html

The problems seems to be when one thread is busy-waiting on global volatile flag,
and another thread sleeps for a given time, and then sets the flag. We can deadlock on
safepoint rendezvous, when "sleeper" thread had reached the safepoint, but "worker" is not
reaching safepoint, blocking progress.

Reproducible test:
 http://cr.openjdk.java.net/~shade/8064749/Stuck.java

$ java Stuck
(exits after 5 seconds)

$ java -XX:-UseCompilerSafepoints Stuck
(waits indefinitely, does not react to jstack -F, SIGINT, etc.)

$ java -XX:-UseCompilerSafepoints -XX:+SafepointTimeout Stuck
("Worker" thread is stuck)

# SafepointSynchronize::begin: Timeout detected:
# SafepointSynchronize::begin: Timed out while spinning to reach a safepoint.
# SafepointSynchronize::begin: Threads which did not reach the safepoint:
# "Worker" #9 prio=5 os_prio=0 tid=0x00007feddc0db800 nid=0x19f0 runnable [0x0000000000000000]
   java.lang.Thread.State: RUNNABLE

# SafepointSynchronize::begin: (End of list)